Comparing Orchestration in Claude, ChatGPT, and Codex

Services that all appear to “use multiple agents” expose very different levels of control. A chat interface usually lets a person specify the desired result and research boundaries. A coding agent may also expose each child agent’s model, permissions, and configuration files. Understanding that distinction prevents a design from depending on controls the chosen service does not provide.

Each service exposes a different orchestration boundary

The orchestration boundary is how much control a person has over work decomposition, child roles, models, permissions, and execution environments. Claude Chat and Cowork are driven mainly through natural-language requests. Claude Code and Codex can store roles and constraints in repository files for reuse by a team.

What this page covers

• Explain the orchestration differences among Claude Chat, Cowork, and Claude Code
• Decide how subagent use differs between ChatGPT Work and Codex
• Choose a surface for research, artifact creation, or software development
• Separate official product capabilities from operating patterns designed by a user

Do not compare products only by whether they have subagents. Check whether roles and models can be fixed, whether work history is visible, and whether the agents can edit files.

Scroll horizontally→

Surface	Primary work	Orchestration unit	Controls that are easy to specify	Main constraint
Claude Chat / Research	Questions, analysis, research	One conversation and Research’s internal investigation	Question, research scope, connected sources, output format	Users do not manage a configuration file for each child
Claude Cowork	Documents, spreadsheets, file organization, long-running knowledge work	Plan, subtasks, parallel work, artifacts	Completion criteria, target files, tools, approval conditions	Focuses on outcomes and permissions more than per-child model configuration
Claude Code	Software development and repository work	Main conversation, subagents, Agent Teams	Agent definition, model, effort, tools, permissions, worktree	Requires control of concurrent-edit conflicts and token growth
ChatGPT	Conversation and general file-based work	Conversation, selected capabilities, Skills	Goal, input, output format, Skill to use	Subagent availability depends on the account and product surface
ChatGPT Work	Long-running work and multi-perspective research or creation	Parent work plus parallel subagents	Decomposition axes, wait conditions, synthesis format	Runs in a hosted environment distinct from local Codex child-agent configuration
Codex	Software development and local or cloud work	Main thread, subagent threads, Skills	Model, reasoning effort, sandbox, concurrency, agent instructions	Each worker execution also consumes tokens and tool use

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Claude Chat expands research dynamically with Research

Standard Claude Chat centers on conversational questions and analysis. For investigations that cross several sources, enabling Research lets Claude decide what to investigate next, run repeated searches, and produce an answer with citations.[1]

Anthropic describes the underlying research system as an orchestrator-worker pattern. A lead agent creates a plan, runs multiple subagents in parallel, and synthesizes their findings.[2]

The user defines research boundaries and completion criteria, not an internal agent directory. A request might look like this:

Use Research to investigate current generative AI governance in Japan.
Examine regulation, industry guidance, and public policies from major companies as separate angles.
Prioritize primary sources. Finish by separating common themes, differences, and unresolved questions.

The request makes the research axes independent and defines the synthesis format. It does not configure each subagent’s model or tools as Claude Code does.

Claude Cowork carries knowledge work from planning to artifacts

Claude Cowork brings the agentic architecture behind Claude Code to knowledge work without requiring terminal operation. It analyzes a request, creates a plan, divides the work into subtasks when useful, and can run multiple tasks in parallel.[3] Its outputs can include completed documents, spreadsheets, presentations, and organized files.

In Cowork, define the input, artifacts, permissions, and review points more carefully than the child-agent topology.

Create a quarterly report from customer interview notes in the connected folder.
The following work may run in parallel:
1. Classify customer problems
2. Count feature requests
3. Extract important quotations

Deliver a Markdown report with source links and a spreadsheet containing the counts.
Do not delete files or send anything externally. Ask before taking either action.

Naming both parallel work and prohibited actions makes safety boundaries part of the orchestration design, not just speed.

Claude Code separates subagents from Agent Teams

A Claude Code subagent performs bounded work in its own context and returns a summary to the main conversation. Markdown definitions in .claude/agents/ can set the role, model, effort, tools, permissions, Skills, and worktree isolation.[4]

Agent Teams, by contrast, is an experimental feature in which a team lead and several teammates coordinate through a shared task list and messages. It offers more coordination than subagents, but each teammate runs as an independent Claude Code instance and therefore increases token use.[5]

The key question is whether children must communicate with one another.

• Use a subagent when the parent only needs the result of research, log analysis, or a bounded review
• Use Agent Teams when several owners must exchange intermediate findings and claim tasks
• Use one session when work has many dependencies or must edit the same file in sequence

See Parent-Child Agent Design in Claude Code for a concrete directory and configuration.

Distinguish standard ChatGPT from ChatGPT Work

ChatGPT combines conversations, Projects, Skills, files, and tools. The official subagent documentation describes parallel agent activity and result views for eligible ChatGPT Work accounts.[6] Do not generalize that every ChatGPT account exposes the same subagent behavior; check the capabilities available in the account and interface being used.

A multi-perspective review request in ChatGPT Work might be:

Review this proposal with parallel subagents.
Assign one agent each to legal wording, numerical consistency, and reader comprehension.
Wait for all three agents, remove duplicate findings, and synthesize the result by severity.
Do not edit the document; return findings and evidence only.

The four explicit controls are decomposition, waiting, synthesis, and non-modification. ChatGPT Work selects a model and intelligence level through the conversation surface; it does not assume the same TOML configuration used by local Codex.[6]

Codex fixes roles through agent threads and configuration files

In a Codex subagent workflow, a parent thread delegates independent work to child agent threads and integrates their results when they finish. In Codex CLI, IDE integrations, and the ChatGPT desktop app, a user can open child threads to inspect progress and results.[6]

Project-level TOML files in .codex/agents/ can configure model, model_reasoning_effort, sandbox_mode, Skills, and MCP connections by role.[6] Reusable procedures belong in .agents/skills/<skill-name>/SKILL.md, so their full instructions load only when needed.[7]

A Codex request can name roles and dependencies directly:

The parent agent must settle the specification and change boundary.
Delegate impact analysis to explorer and missing-test analysis to test_analyst in parallel.
Wait for both results before assigning one minimal implementation to worker.
Finish with reviewer inspecting only the diff.

See Parent-Child Agent Design in Codex for concrete TOML and Skill designs.

Choose a surface from the shape of the work

First decide whether the task ends in an answer, research report, artifact, or repository change. Then ask whether it contains independent perspectives, whether child agents need to communicate, and whether models and permissions must be fixed by the user.

flowchart TD
    A["What result is needed?"] --> B{"Will files or code change?"}
    B -->|No| C{"Is this long-running artifact creation?"}
    C -->|No| D["Claude Chat / ChatGPT"]
    C -->|Yes| E["Claude Cowork / ChatGPT Work"]
    B -->|Yes| F{"Is repository-level configuration required?"}
    F -->|Yes| G["Claude Code / Codex"]
    F -->|No| E

Product internals are not necessarily equivalent. Treat only published settings as user-controlled capabilities, and describe internal model topology only when official material documents it.

References

1. Anthropic, Use research on Claude, 2026-06-02
2. Anthropic, How we built our multi-agent research system, 2025-06-13
3. Anthropic, Get started with Claude Cowork
4. Anthropic, Create custom subagents
5. Anthropic, Orchestrate teams of Claude Code sessions
6. OpenAI, Subagents
7. OpenAI, Build skills

For details of the latest releases and updates, refer to the official websites and documentation.

Related links

• Orchestration Patterns
• AI Agent Framework Comparison
• What Are Subagents?